4.8.P Current achievements and future developments of a novel AI based visual monitoring of beehives in ecotoxicology and for the monitoring of ...

Page created by Michele Hayes
 
CONTINUE READING
Hazards of pesticides to bees - 14th international symposium of the ICP-PR Bee protection group, October 23 – 25 2019, Bern (Switzerland)
                                                              Abstracts: Poster

4.8.P Current achievements and future developments of a novel AI based visual
monitoring of beehives in ecotoxicology and for the monitoring of landscape
structures
Frederic Tausch, Katharina Schmidt, Matthias Diehl
apic.ai GmbH Rintheimerstraße 31-33, 76131 Karlsruhe, Germany,
E-Mail: frederic.tausch@apic.ai
DOI 10.5073/jka.2020.465.064

Abstract
Honey bees are valuable bioindicators. As such, they hold a vast potential to help shed light on the extent and
interdependencies of factors influencing the decline in the number of insects. However, to date this potential
has not yet been fully leveraged, as the production of reliable data requires large-scale study designs, which are
very labour intensive and therefore costly.
A novel Artificial Intelligence (AI) based visual monitoring system could enable the partial automatization of data
collection on activity, forager loss and impairment of the central nervous system. The possibility to extract
features from image data could prospectively also allow an assessment of pollen intake and a differentiation of
dead bees, drones and worker bees as well as other insects such as wasps or hornets.
The technology was validated in different studies with regards to its scalability and its ability to extract motion
and feature related information.
The prospective possibilities were analyzed regarding their potential to enable advances both within
ecotoxicological research and the monitoring of pollinator habitats.
Keywords: Artificial Intelligence, Oomen feeding, colony development, novel method, hive monitoring, bee
counter, honey bees, bioindicators, ecotoxicology, activity assessment

Introduction
Honey bee colonies can act as detectors for harmful substances by either signaling the existence of
toxic molecules through high mortality rates or by accumulating residues for not acutely lethal
substances (e.g., of heavy metals, fungicides and herbicides) in pollen, nectar or larvae (Celli, 1983;
Porrini et al., 2002). They were first used as bio-indicators to monitor environmental quality in 1935
(Crane, 1984). The detection of pesticide use is one of the research fields in which bee monitoring
has since been applied (Atkins et al., 1981; Celli, 1983; Mayer and Lunden, 1986; Mayer et al., 1987;
Celli et al., 1988; Celli and Porrini, 1991; Celli et al., 1991; Porrini et al., 1996). With about a quarter of
its inhabitants being active foragers, the condition of a colony mirrors the state of its habitat. Among
the requisites which make the colonies especially suitable environmental indicators are that they
can be easily held by beekeepers, that their foragers cover large areas and that they collect samples
like pollen or nectar out of self-interest. (Celli and Maccagnani, 2003).
Honey bee colony development depends on many factors such as but not limited to queen age,
nutrition, colony strength, pathogens and parasites as well as regional particularities. Therefore,
large sample sizes are necessary in order to generate objective insights into causal relationships of
hazards towards honey bees. Within the German bee project, which was aimed at understanding
the causes of honey bee colony collapses, more than 1.200 bee hives in 125 places across the
country were monitored between 2004 and 2009. The study brought many correlations to light but
also left some questions open. The authors presume that a study design suitable to record sublethal
or chronic effects might reveal a negative effect of pesticides regarding colony collapse which they
were not able to detect. (Genersch et al., 2010).
Because large scale studies using bees as bioindicators are very time and labour intensive, their
number is still quite small. In 1978, Giordani et al. were able to demonstrate the highly toxic effect
of the chlorinated hydrocarbon insecticide Endosulfan. Yet, many years and several studies were
needed to provide enough evidence to change a limitation of the use of the substance. Later, in a
large-scale monitoring project in northern Italy, bee mortality was recorded for several hundred

124                                                                                                             Julius-Kühn-Archiv, 465, 2020
Hazards of pesticides to bees - 14th international symposium of the ICP-PR Bee protection group, October 23 – 25 2019, Bern (Switzerland)
                                                              Abstracts: Poster
hives under both high and low chemical pressure from farming. By analyzing dead bees from hives
with especially high numbers of casualties it was possible to identify the molecules responsible for
76% of registered mass-deaths. Nevertheless, one shortcoming of the design the authors
mentioned was that the number of dead bees collected was only a conservative estimate as the
losses caused by lethal doses in the field could not be recorded. (Celli and Maccagnani, 2003).
These works demonstrate the potential of bee monitoring in various fields from pesticide regulation
to general advances in research regarding bee health. However, they are pioneer projects and not
representative for the way research is typically conducted. To date, factor analysis and preventive
activities are built mostly on snapshot data from a small number of hives which can be collected
more economically. The use of technology could help decrease the labour intensity and therefore
the cost of such projects. A few systems, based on different technologies has recently been
developed, yet still exhibit shortcomings.
There are counting systems, which try to quantify the incoming and outgoing bees at the entrance
(e.g., BeeCheck with capacitive detection) (Gombert et al., 2019). Due to their design, the counting
systems only record the pollinators within a short distance. The information content of their sensory
raw data is considerably reduced for evaluation with an imaging method. In complex situations,
such as bees running on top of each other or group formation they are therefore prone to
measurement inaccuracies and would consequently not be suitable for a robust assessment of
mortality. With a visual system it is possible to follow each individual animal over a sequence of
images. First scientific works could already present prototype systems, which used a camera system
at the hive entrance to determine the parasite infestation (Schurischuster et al., 2018).
From 2017 to 2020, the EU-funded project IoBee aims to identify global changes in bee populations
through the networking of data from bee colonies. The data collection within the project is
supported by technical sensor systems. However, these partly specialized, partly integrated sensor
systems detect only temperature, humidity, sound and weight to determine the health of individual
hives. They are not designed to collect information about activity, forager loss or foraging intensity.
AI has a high potential to contribute to data in areas, which can only be collected from images
(Bozek et al. 2018). Because Neural Networks, like all software are scalable and once trained able to
extract results very precisely, their use could vastly improve the quantity and quality of information
which can be gathered from the monitoring of bee hives.

Materials and Methods

Fig. 1 Visual monitoring device in front of the hive entrence
An AI-based visual monitoring technology for bee hives has been developed (Fig. 1), in order to
create a cost-efficient way to autonomously collect data on a larger scale with high quality. It
combines hardware and software components. A camera device is attached to the hive entrance to
record all bees entering and leaving. As there are usually no energy and network connections
available at bee hives during field studies, the device was equipped with a solar panel. A UMTS-

Julius-Kühn-Archiv, 465, 2020                                                                                                               125
Hazards of pesticides to bees - 14th international symposium of the ICP-PR Bee protection group, October 23 – 25 2019, Bern (Switzerland)
                                                              Abstracts: Poster
connection makes sure the software on the device can be accessed and updated remotely. Different
software modules subsequently analyze the image data both on-device and with the help of cloud
computing resources. Deep learning algorithms are used which have been trained through
exposure to selected exemplary data sets. These methods of AI enable the collection of objective
data and still allow a verification of the results at a later time, because image segments with
individual bees and video sections are simultaneously and sequentially archived in a database. The
technology can be applied to generate both motion and feature related insights.

Fig. 2 Visualisation of the motion Analysis (A) and visualization of the feature analysis (B)

Motion-related analysis:
The ingress and egress of bees is captured on camera at the hive entrance. Neural Networks detect
each bee and track its movement while it passes through the camera’s field of view, which measures
about 145 x 108 mm² (Fig. 2, A). Subsequently, different aspects can be analyzed:
 Level of activity derived from number of incomings and outgoings
 Loss as the proportion of outgoings which do not return
 Motion patterns of individual bees within the cameras window frame

Feature-related analysis:
Cropped single-entity images of bees entering and leaving their hive are uploaded into a cloud (Fig.
2, B). The collected data are processed by a multi modal neuronal network to perform the following
tasks:
 Recognition of pollen on bee’s legs;
 Detection of whether a bee is dead or alive;
 Differentiation of drones and workers; and,
 Differentiation of other genera like wasps or hornets.
In 2019, the monitoring system was tested in two test settings to validate different aspects of both
the hardware and software components within the technology.
The practical scalability of the technology was tested as a precondition to apply it under realistic
circumstances. The camera devices were attached to a total of 33 colonies in 14 locations in and
around the city of Karlsruhe, at hives of local bee keepers, public institutions or companies.
In a different setting, bee activity was measured to detect changes in flight activity following
neonicotinoid exposure. Within an Oomen feeding study in an agricultural area, the applicability of
using the visual bee’s activity in pollinator risk assessment was evaluated. The study assessed the
impact of the feeding with a neonicotinoid on daily activity and colony development. Eight hives
were monitored, of which four were fed with 500 g of a sugar solution, including a concentration of
200 µg imidacloprid/ kg of sugar solution, for ten consecutive days. The control group was fed the
same amount of sugar solution during that exposure period (Gonsior et al., 2019).

126                                                                                                             Julius-Kühn-Archiv, 465, 2020
Hazards of pesticides to bees - 14th international symposium of the ICP-PR Bee protection group, October 23 – 25 2019, Bern (Switzerland)
                                                              Abstracts: Poster
In a separate proof-of-concept study, the possibilities of machine learning algorithms were explored
to perform localization, classification and pose estimation tasks on video recordings (Marstaller et.
al., 2019).

Results
The results of the 2019 studies show first positive achievements for future application in the fields
of ecotoxicology and for the monitoring of landscapes.
For both usecases, it is essential to build a scalable and failsafe system. Special methods for failure
detection were developed, to ensure the uptime of the camera devices. Cloud monitoring alerts
were set up for notification in case of failures to reduce downtime. Adding these mechanisms on
different system levels made the devices independent and self-sufficient. This will make it possible
in the future to continuously monitor large areas and remote locations.
Regarding the results of the Oomen study, Fig. 3 shows the change in activity per hive during the
exposure period. The negative values represent bees leaving their hive while positive values
represent bees returning. Values are plotted as the sum of bees per hour. During the exposure
period, from the third day of exposure onwards the hives fed with the neonicotinoid displayed a
significantly decreased level of activity. It has thus been demonstrated, the technology used can
collect relevant parameters for ecotoxicological reseasrch, which could not be assessed before.

Fig. 3 Activity at the hive entrance following contact with imidacloprid (Gonsior et al., 2019)
The detection of whether individual animals were drones, worker bees or different genuses like
wasps and whether they were dead or alive could be achieved with the use of a cloud based multi
network. It was also possible to detect whether bees carried pollen or not with a certain probability.
Additionally, information on the bee’s pose could be recognized, including detailed information on
the movement. This pilot work suggests that generating further insights beyond the activity of bees
is feasible, which amplifies the advantages of using bees as bio-indicators.

Conclusion
In first tests, AI-based visual monitoring of bee hives has shown great potential to reliably capture
and analyze honey bee’s motion and features. With the help of Neural Networks honey bees can be
used as bio-indicators in new ways. Information about environmental factors can be collected,
which has not yet been accessible.
A test-field with a networked system of prototypes for real-time analysis is in place in Karlsruhe for
further testing. Algorithms are currently developed to be deployed in 2020 using this infrastructure.
Data on activity, forager loss and food availability will be systematically collected. For this purpose,
beehives will be monitored in rural and urban landscapes all over Baden-Württemberg. The aim of
the project is to identify local and seasonal problems and evaluate landscape suitability as a habitat
for insects. As the camera devices in the field can be updated remotely, software improvements can
be distributed whenever algorithms are improved. The accuracy of the algorithms depends on the
training data within the database. Therefore, it can be improved either by providing additional data

Julius-Kühn-Archiv, 465, 2020                                                                                                               127
Hazards of pesticides to bees - 14th international symposium of the ICP-PR Bee protection group, October 23 – 25 2019, Bern (Switzerland)
                                                              Abstracts: Poster
sets or by improving the existing input data through data annotation and quality assurance by
experts.
It has been proven feasible to set up a connected network of bee hives to remotely monitor hive
activity. Tests of the scalability of the technology give reason to believe that it can be used to
constantly and simultaneously numerous hives. By allowing the monitoring of large numbers of
colonies with minimal human labour, AI based visual monitoring of bee hives could therefore
become a valuable tool in ecotoxicological field studies. It could in the future enable significantly
larger study designs and therefore facilitate more reliable results. Once this it possible, the
technology could be used to classify the magnitude of the detrimental effects of plant protection
products as well as other environmental hazards on colonies.
In a collaborative study with Eurofins Agroscience Services Ecotox it has already been possible to
quantitatively track bee activities. Improvements are currently in progress to achieve a level of
accuracy at which an accurate determination of the precise loss of foragers can be derived from the
quantitative ingress and egress activity. It could thus become possible for the first time to generate
precise data of the loss of bees which died outside their hive because of lethal environmenta effects.
The data stream would be constant and could be collected without human action or assessment
bias.
In addition to information about the level of activity and the precise loss of foragers, the motion
profiles of each bee could be analyzed as an indicator of sub-lethal or chronic effects. Changes in
motion patterns could indicate an impairment of the central nervous system or health related issues
of colonies such as the deformed wing virus.
Furthermore, a quantitative assessment of pollen intake could be used to assess the extent of
foraging activity. Low pollen intake at a number of neighbouring hives might indicate a temporal
shortage in food supply within the respective region. This information could be utilized to
implement targeted measures like the cultivation of plants which bloom during that period. As a
result, the food situation would be improved not only for the honey bees but also for other
pollinators. In this case the honey bee colonies could serve as a bio-indicator to identify
environments in which there would not be sufficient food available to feed pollinators all year long.
If merged with further information, such data could also contribute to ecological impact statements.

References
ATKINS E. L., KELLUM D., ATKINS K. W., 1981: REDUCING PESTICIDES HAZARD TO HONEY BEES: MORTALITY PREDICTION TECHNIQUES AND INTEGRATED
     MANAGEMENT STRATEGIES. DIVISION OF AGRICULTURAL SCIENCES, UNIVERSITY OF CALIFORNIA, LEAF. 2883, 22 PP.
BOZEK, K., HEBERT, L., MIKHEYEV, A.S. AND STEPHENS, G.J., 2018: TOWARDS DENSE OBJECT TRACKING IN A 2D HONEYBEE HIVE. PROCEEDINGS OF THE
     IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, SALT LAKE CITY, 4185-4193.
CELLI, G., 1983: L’APE COME INSETTO TEST DELLA SALUTE DI UN TERRITORIO. ATTI XIII CONGRESSO NAZIONALE ITALIANO ENTOMOLOGIA, SESTRIERE –
     TORINO. 637-644.
CELLI, G., PORRINI, C., 1991: L’APE, UN EFFICACE BIOINDICATORE DEI PESTICIDI. LE SCIENZE 274, 42-54.
CELLI, G., PORRINI, C., BALDI, M., GHIGLI, E., 1991: PESTICIDES IN FERRARA PROVINCE: TWO YEARS’ MONITORING WITH HONEY BEES (1987-1988).
     ETHOLOGY, ECOLOGY AND EVOLUTION SPECIAL ISSUE 1, 111-115.
CELLI, G., MACCAGNANI, B., 2003:, HONEY BEES AS BIOINDICATORS OF ENVIRONMENTAL POLLUTION, BULLETIN OF INSECTOLOGY 56, 137-139.
CELLI, G., PORRINI, C., TIRAFERRI, S., 1988: IL PROBLEMA DEGLI APICIDI IN RAPPORTO AI PRINCIPI ATTIVI RESPONSABILI (1983-1986). ATTI GIORNATE
     FITOPATOLOGICHE, LECCE, 257-268.
CRANE, E., 1984: BEES, HONEY AND POLLEN AS INDICATORS OF METALS IN THE ENVIRONMENT. BEE WORLD 55. 47-49.
GENERSCH, E., VON DER OHE, W., KAATZ, H. ET AL., 2010: THE GERMAN BEE MONITORING PROJECT: A LONG TERM STUDY TO UNDERSTAND PERIODICALLY
      HIGH WINTER LOSSES OF HONEY BEE COLONIES. APIDOLOGIE 41, 332–352.
GOMBERT, A., BERMING, S.,FROMMBERGER, M. PISTORIUS, J., 2019: HOW CAN THE BEECHECK BE USED IN RISK ASSESSMENT TO QUANTIFY A PESTICIDES
      IMPACT ON BEES? 11TH JULIUS KÜHN INSTITUTE YOUNG SCIENTISTS MEETING, 2018, BRAUNSCHWEIG
GONSIOR, G., TAUSCH, F., SCHMIDT, K.L., KNÄBE, S., 2019: IMPACT OF AN OOMEN FEEDING WITH A NEONICOTINOID ON DAILY ACTIVITY AND COLONY
     DEVELOPMENT OF HONEYBEES ASSESSED WITH AN AI BASED MONITORING DEVICE. PROCEEDINGS OF THE 14TH INTERNATIONAL SYMPOSIUM: HAZARDS
     OF PESTICIDES TO BEES, BERN.
MARSTALLER, J., TAUSCH, F., STOCK, S., 2019: DEEPBEES-BUILDING AND SCALING CONVOLUTIONAL NEURONAL NETS FOR FAST AND LARGE-SCALE
     VISUAL MONITORING OF BEE HIVES. PROCEEDINGS OF THE IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV) 2019, SEOUL.
MAYER, D. F., JOHANSEN, C. A., LUNDEN, J. D., RATHBONE, L., 1987: BEE HAZARD OF INSECTICIDES COMBINED WITH CHEMICAL STICKERS. AMERICAN BEE
     JOURNAL 127 (7), 493-495.

128                                                                                                             Julius-Kühn-Archiv, 465, 2020
Hazards of pesticides to bees - 14th international symposium of the ICP-PR Bee protection group, October 23 – 25 2019, Bern (Switzerland)
                                                              Abstracts: Poster
MAYER, D. F., LUNDEN, J. D., 1986: TOXICITY OF FUNGICIDES AND AN ACARICIDE TO HONEY BEES (HYMENOPTERA: APIDAE) AND THEIR EFFECTS ON BEE
    FORAGING BEHAVIOR AND POLLEN VIABILITY ON BLOOMING APPLES AND PEARS. ENVIRONMENTAL ENTOMOLOGY 15: 1047-1049.
OOMEN P.A., DE RUIJTER, A., VAN DER STEEN, J., 1992: METHOD FOR HONEYBEE BROOD FEEDING TESTS WITH INSECT GROWTH-REGULATING INSECTICIDES.
    - EPPO BULLETIN 22, 613 - 616.
PORRINI, C., COLOMBO, V., CELLI, G., 1996: THE HONEY BEE (APIS MELLIFERA L.) AS PESTICIDE BIOINDICATOR. EVALUATION OF THE DEGREE OF POLLUTION
    BY MEANS OF ENVIRONMENTAL HAZARD INDEXES. PROCEEDINGS XX INTERNATIONAL CONGRESS OF ENTOMOLOGY, FIRENZE, 444.
PORRINI, C., GHINI, S., GIROTTI, S., SABATINI, A. G., GATTAVECCHIA, E., CELLI, G., 2002: USE OF HONEY BEES AS BIOINDICATORS OF ENVIRONMENTAL
    POLLUTION IN ITALY. HONEY BEES: ESTIMATING THE ENVIRONMENTAL IMPACT OF CHEMICALS, LONDON AND NEW YORK, 186-247.
SCHURISCHUSTER, S., REMESEIRO, B., RADEVA, P., KAMPEL, M., 2018: A PRELIMINARY STUDY OF IMAGE ANALYSIS FOR PARASITE DETECTION ON HONEY
    BEES. IMAGE ANALYSIS AND RECOGNITION, 465-473.

4.9.P Pollinator monitoring in agroecosystems – general methods for evaluations
in field studies
Julian Fricke, Olaf Klein, Silvio Knäbe
Eurofins Agroscience Services Ecotox GmbH, Eutinger Str. 24, 75223 Niefern-Öschelbronn, Germany
E-Mail: JulianFricke@eurofins.com, OlafKlein@eurofins.com, SilvioKnaebe@eurofins.com
DOI 10.5073/jka.2020.465.057

Abstract
Extensive knowledge of the occurrence, condition and population changes of wild bee communities in
agroecosystems is important. The knowledge is needed to understand the complexity of potential exposure
routes to plant protection products in specific crops and agricultural scenarios or to evaluate possible impacts
of treatments at a landscape scale taking into account other influencing parameters like the cultivation system
or management practices.
Keywords: pollinator, monitoring, solitary bees, risk assessment, experimental design, non-Apis

Introduction
Pollinator monitoring studies are performed under field conditions. They focus on native bee
communities occurring in agroecosystems and can be useful to make spatial and temporal
comparisons in a multifaceted context to allow conclusions regarding the causes of community and
development changes. They can therefore provide an important database for the design and
evaluation of strategies and concrete measures to support and conserve wild bee communities in
agroecosystems.
Generally, the abundance and species richness of naturally occurring pollinators in a crop and
adjacent field margins will be investigated. For crops considered to be not attractive as foraging and
nesting habitats for honey bees, wild bees and other pollinators, the comparison of in-field and off-
crop abundance and richness can help to understand if pollinators are exposed to plant protection
products or not. This might include temporal as well as spatial differences (timing of monitoring and
placement of monitoring within the field and landscape).

Materials and Methods
To evaluate a wide range of pollinator species occurring in a specific crop, several methods are
available. We recommend using a combination of different types of sampling techniques: non-
selective and selective, because wild bees are often highly specialized in their floral choices, nesting
behavior and phenology e.g., so that their populations can undergo strong spatio-temporal
variations. For the non-selective methods two different types of traps might be used in combination:
Vane traps and Bee bowls (pan traps). These traps can be installed at different locations (i.e., in the
centre of the fields; at the borders of the fields; and, outside in the adjacent field margin) and
different heights adapted to the type of the crop which is investigated. As selective method, sweep
netting and observation can be used via standardized or variable transect walks in a defined
distance and time interval or at fixed locations.

Julius-Kühn-Archiv, 465, 2020                                                                                                               129
You can also read